In an example, an image classification system is disclosed. The image classification system modifies an image having a first height and a first width to be input to a convolutional neural network for image classification. The image classification system includes an image resizing module that is configured to resize the image so that the resized image comprises a second height and a second width. An aspect ratio of the resized image corresponds to an aspect ratio of the image having the first height and the first width. The image classification system also includes an alignment module that is configured to modify pixels of a feature map corresponding to the resized image based upon a comparison of a desired feature map size and an actual feature map size.
Legal claims defining the scope of protection, as filed with the USPTO.
1. An image classification system for modifying an image having a first height and a first width to be input to a convolutional neural network for image classification, the image classification system comprising: an image resizing module that is configured to resize the image so that the resized image comprises a second height and a second width, wherein an aspect ratio of the resized image corresponds to an aspect ratio of the image having the first height and the first width; and an alignment module that is configured to modify pixels of a feature map corresponding to the resized image based upon a comparison of a desired feature map size and an actual feature map size; wherein the second width is equal to int(a 1 *√{square root over (r)}), where int is an integer operation, a 1 is at least one of a desired number of pixels of a width of the resized image or a desired number of pixels of a height of the resized image and r is the aspect ratio of the image having the first height and the first width.
2. The image classification system as recited in claim 1 , wherein the second height is equal to int ( a 1 r ) , where int is an integer operation, a 1 is at least one of a desired number of pixels of a width of the resized image or a desired number of pixels of a height of the resized image and r is the aspect ratio of the image having the first height and the first width.
3. The image classification system as recited in claim 1 , wherein the alignment module is further configured to at least one of remove the pixels from the feature map or add the pixels to the feature map based upon the comparison.
4. The image classification system as recited in claim 3 , wherein the alignment module is further configured to remove the pixels from the feature map when the comparison indicates the actual feature map size is greater than the desired feature map size.
5. The image classification system as recited in claim 4 , wherein the alignment module is further configured to add the pixels to the feature map when the comparison indicates the actual feature map size is less than the desired feature map size.
6. The image classification system as recited in claim 1 , further comprising a convolution module that is configured to receive the resized image and output the feature map based upon the resized image.
7. The image classification system as recited in claim 6 , wherein the convolution module is further configured to apply one or more convolution operations to the resized image.
8. The image classification system as recited in claim 1 , wherein the aspect ratio of the resized image approximately equals the aspect ratio of the image having the first height and the first width.
9. The image classification system as recited in claim 1 , further comprising a predictor module that is configured to generate a prediction based upon the feature map.
10. A method for modifying an image having a first height and a first width to be input to a neural network for image classification, the method comprising: resizing the image so that the resized image comprises a second height and a second width, wherein an aspect ratio of the resized image corresponds to an aspect ratio of the image having the first height and the first width; and modifying pixels of a feature map corresponding to the resized image based upon a comparison of a desired feature map size and an actual feature map size; wherein the second width is equal to int(a 1 *√{square root over (r)}), where int is an integer operation, a 1 is at least one of a desired number of pixels of a width of the resized image or a desired number of pixels of a height of the resized image and r is the aspect ratio of the image having the first height and the first width.
11. The method as recited in claim 10 , wherein the second height is equal to int ( a 1 r ) , where int is an integer operation, a 1 is at least one of a desired number of pixels of a width of the resized image or a desired number of pixels of a height of the resized image and r is the aspect ratio of the image having the first height and the first width.
12. The method as recited in claim 10 , further comprising at least one of removing the pixels from the feature map or adding the pixels to the feature map based upon the comparison.
13. The method as recited in claim 12 , further comprising removing the pixels from the feature map when the comparison indicates the actual feature map size is greater than the desired feature map size.
14. The method as recited in claim 13 , further comprising adding the pixels to the feature map when the comparison indicates the actual feature map size is less than the desired feature map size.
15. The method as recited in claim 10 , further comprising receiving the resized image and outputting the feature map based upon the resized image.
16. The method as recited in claim 15 , further comprising applying one or more convolution operations to the resized image.
17. The method as recited in claim 10 , wherein the aspect ratio of the resized image approximately equals the aspect ratio of the image having the first height and the first width.
18. The method as recited in claim 10 , further comprising generating a prediction based upon the feature map.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
August 23, 2018
July 21, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.